From an IR perspective, the question is where and how KBIR can be of benefit. One possibility is in the interface. Natural language processing (NLP) methods can be used so that the user enters natural language statements that are ``understood'', and receives natural language responses (e.g., explanations, questions needed for clarification, statements regarding progress of the session, etc.). The user may be ``understood'' or modeled, and that may continue across sessions so that the system adapts to the user needs and characteristics.
A second possibility is that the system actually ``understands'' the users' interest statements and/or the documents. This can result from NLP, and lead to useful knowledge representations of the interest statements and documents.
Third, the system may guide the user, through dialog and possibly through a machine learning scenario, to develop and refine problem descriptions and ultimately query representations. In some cases this may simply involve working with an existing back-end retrieval system, and forming the best possible Boolean query. In other cases, the back-end system may be more flexible. In either case, query expansion and term disambiguation are important issues.
Fourth, if the above occur, the system may use AI methods to match the representations, yielding some type of ranking or selection that considers the uncertainty and the statistical data that are available. Such matching may involve logical or statistical (e.g., Bayesian) inference, or (conceptual) graph matching operations. KBIR methods can also be applied to help increase parallelism (e.g., with blackboards), or to improve the efficiency of these operations, where rules or heuristics or AI search techniques help prune the space.
More details on these approaches are given in terms of an AI perspective on IR operations.